Defect Detection of Fabric Printing with Differential Attention

Abstract

Fabric defect detection is a major quality control process in the textile industry. However, compared with common detection tasks, printed fabric defect detection has several difficulties as follows: printed fabric pattern texture is complex; defects are of numerous types and vary greatly in size and shape; most of the defects are extremely small.Aiming at these problems and difficulties in real scenes, this paper firstly adopts the idea of template matching in change detection to eliminate the background pattern texture features of target images by using template image features.Aiming at the problem of large scale difference of defects, this paper proposed a better modeling ability of Ratio DCN on the basis of deformable convolutional network for defects with different scales, especially the abnormal Ratio of long and short edges. To solve the problem of a large number of small target defects, this paper uses the template image features as an auxiliary structure, and fuses the high-level and low-level features of the detected and template images to improve the detection rate of small target defects.In addition, the Diff Attention structure was also proposed in this paper.Based on the differential features of the target image and template image, the model's Attention was more focused on the region where the defect was located, which could strengthen the feature extraction ability of the model for the defect.DPDAN's good generalization is verified on both its own data set and public data set.